Reproduce “Machine Learning for Analyzing Atomic Force Microscopy (AFM) Images Generated from Polymer Blends” on OpenResearch

About this paper

Aanish Paruchuri, Yunfei Wang, Xiaodan Gu, Arthi Jayaraman

An automated, unsupervised machine learning workflow analyzes atomic force microscopy (AFM) images of polymer blends, enabling objective segmentation of polymer domains and quantification of their size distributions. The optimal workflow, leveraging Discrete Fourier Transform and K-means clustering, achieved an accuracy of 0.81 and a Dice coefficient of 0.85 for domain segmentation.

Key insights

• Frequency-domain transforms (DFT, DCT) proved more effective for unsupervised AFM image segmentation than a pre-trained deep learning model (ResNet50) for this specific application.
• Variance emerged as the most distinct and powerful statistical feature for capturing textural differences between polymer domains from the transform outputs.
• Deep learning models, while powerful, may not be optimal for all scientific image analysis tasks, especially when specific domain features or the lack of labeled data necessitate simpler, more tailored approaches.

Summary by alphaXiv.